Presented at the Neonatal Society 2015 Summer Meeting.

Blesa M¹, Serag AF¹, Wilkinson AG², Moore EJ¹, Pataky R¹, Sparrow SA¹, Macnaught G³, Semple SI³, Bastin M⁴, Boardman JP^1,4

¹ MRC Centre for Reproductive Health, University of Edinburgh, UK
² Department of Radiology, Royal Hospital for Sick Children, Edinburgh, UK
³ Clinical Research Imaging Centre, University of Edinburgh, UK
⁴ Centre for Clinical Brain Sciences, University of Edinburgh, UK

Background: Labelled MRI atlases can be used to: calculate the volume and infer tissue integrity of anatomic regions of interest (ROIs); capture population diversity in brain structure; localise abnormalities; construct growth curves; and to map regional brain growth in response to injury / treatment. Aim: to construct a neonatal atlas with 107 ROIs operable across MRI modalities.

Methods: Participants: 25 healthy control infants born at term underwent multimodal MRI with ethical approval. Framework: Stepwise longitudinal non-rigid registration (1) of a labelled adult brain (SRI24/TZO), via MRI templates from children at 4.5yrs, 2.5yrs and 3 months, to the neonatal template (2); Propagation of the labelled map from the template to structural MRI data from neonatal cohorts using nonlinear image registration (3); Construction of multimodal atlas using an iterative averaging-based approach (4,5). Label accuracy was assessed by a radiologist (AGW). Accuracy between longitudinal and direct registration from adult to neonatal templates was compared using the Dice coefficient. The study was supported by Theirworld and NHS Research Scotland.

Results: We created a neonatal brain template labelled with 107 anatomic ROIs. Manual editing was required for sub-cortical structures, and 17 regions were manually added (lateral ventricles, brainstem, cerebellum and corpus callosum). The figure illustrates colour coded ROIs in three planes and volume rendered cortical regions (bottom left). The computed transformations for constructing the T1-weigthed atlas were used to create T2-weigthed, fractional anisotropy and mean diffusivity average templates, as well as tissue probability maps (CSF, WM, GM and deep GM). The Dice coefficient was greater for all brain regions after longitudinal versus direct registration between adult and neonatal templates. Average coefficient for all regions with longitudinal registration: 0.726; and with direct registration: 0.421.

Conclusion: We present a neonatal MRI atlas that captures population diversity, has rich anatomic definition (107 ROIs), and can be used to analyse multi-modal MRI scans. We also show that longitudinal registration using spatiotemporal atlases enhances accuracy for modelling brain change across the life-course.

Corresponding author: M.Blesa@sms.ac.ed.uk

References
1. Serag et al. 2012
2: Fonov et al. 2009
3: Avants et al. 2008
4: Guimond et al. 2001
5: Joshi et al. 2002